Investors and forecasters often extrapolate from past returns, but whether this reffects behavioral bias or efficient information processing remains unclear. We address this questionby inferring Chinese mutual fund managers’ market expectations from textual analysis oftheir commentaries and linking them to portfolio choices and performance. Extrapola-tion is state-dependent: it is stronger when growth is above trend and idiosyncratic riskis relatively more important. It is associated with weaker market timing and strongerstock picking, leaving overall performance unchanged. Our findings support a rational-inattention model of expectation formation, in which managers shift scarce attentionbetween aggregate and stock-speciffc information as the relative importance of differentrisks change.

In forecast surveys of aggregate macroeconomic and financial variables, the correlation between forecast errors and forecast revisions is positive at the consensus level, but negative at the individual level. Past literature has interpreted this discrepancy as evidence of underreaction to news at the aggregate level and overreaction at the individual level. In this paper, I challenge this view by arguing that noise in predictive judgment can account for the difference. Using a stylized model, I examine how introducing judgment noise at the individual level changes the interpretation of the correlation coefficients. First, a negative coefficient at the individual level no longer necessarily means overreaction. Second, the coefficient at the consensus level underestimates the degree of underreaction. Using forecast survey data, I provide evidence that judgment noise is large enough to reconcile the difference between the two coefficients. The structural parameter measuring over-/underreaction mainly points to underreaction, regardless of whether the model matches correlation coefficients at the individual or aggregate level.